· ·
·

Hi, I'm Mrinaal! 👋

I work independently on AI research full-time, focused on AI alignment, AI safety, multilingual models and agent evaluation. Along the way, I've worked hands-on with language-model training and post-training, reinforcement learning, mechanistic interpretability and multi-agent systems. I also maintain isindiakeepingup.com, which tracks India's LLM release pace against the US, China, Singapore and South Korea.

Previously, I was a Software Engineer at Ohara, where I owned the core frontend, shipped it as a live Base miniapp and built the integration pipeline that let AI-generated apps run inside Base. Before that, I interned at OK Health, building their embeddable AI chatbot.

I graduated Summa Cum Laude with a B.S. in Data Science from Arizona State University in 2025. During college I won 6+ hackathons, including at Harvard and Yale, and built viral websites like Your-Wrapped, iammusic2024.com, and  World Fan Arena, which together reached 200k+ unique visitors.

at academic or industry AI labs. See my research CV, and if you're working on related problems, email me.

Open Source Projects

Claude Codex Canvas — A Claude Code skill for creating images and animations with Codex.

Codex Claude Crew — Codex delegates to Claude Code, then verifies.

Rumik TTS Fast Inference — Optimizing Rumik TTS on one H100.

Emergent Misalignment — Narrow fine-tuning, broad behavioral shifts.

Cross-Lingual Emergent Misalignment — English fine-tuning, misalignment across languages.

Efficient Qwen — Autoresearch for faster Qwen3.5-4B inference on one GPU.

Parameter Golf — Efficient pretraining experiments for OpenAI’s 16 MB language model challenge.

Qwen3 Sparse Autoencoders — Finding interpretable features in Qwen3-4B and steering it toward cooking instructions.

Kaarya — A minimal iOS task list with a Pomodoro timer.

Duotime — A second timezone in your macOS menu bar.

Cursor Path — A Chrome extension that measures how far your cursor travels.